LLM Representation Learning

Improving LLM reasoning abilities by disentangling its representation extraction abilities from reasoning on these representations. Contains:

1. classification on full dataset of inner states (with or without few-shot examples, with or without CoT)
2. classification on small sequences of inner states

Currently supports MCQA datasets only.

Usage

Overwrite constants.py, and run the numbered python files 1-compute-states.py, etc. in order.